Comments (9)
IIRC it’s a problem with the astra algorithm itself.
I want to say that the forward projector is always pixel based, but the GPU adjoint is ray based? Don’t quote me on that.
from scico.
So if it is due to the inaccuracy of the astra projector, should we increase the tolerance of the test or just disable them for gpu?
from scico.
To clarify: we know there is at least a problem with the ASTRA GPU backprojection algorithm itself. The question here why is the new adjoint test failing with a relative error approaching one. I didn't think the ASTRA backprojection was that inaccurate, but I don't have a good intuition for our new adjoint test.
If you are convinced there is no bug in the interface or test, I learn toward increasing the tolerance rather than disabling the test, because it might allow us to catch regressions in the interface in the future.
from scico.
rel err of 1 is larger than i'd expect; iirc it was something like 10%?
has the astra version changed? it looks like they finally shipped v2 (was on 1.9.9 beta for a looong time)
has the jax/jaxlib version changed? might be worth seeing if something has been modified in the host_callback interface; it is marked as "experimental" and subject to breaking changes
from scico.
Is the adjoint test new? Did it ever pass on GPU for astra?
from scico.
No, it isn't new. It passed with very, very loose tolerances; like a relative error of 20% or something.
Sounds like that's no longer the case (and I have no gpu machine to test)
from scico.
I did pytests on old commits to see where it broke on GPU.
It seems the tests pass on commit d86b51f and fail after commit 656f59d (Improve LinearOperator adjoint tests (#72)).
That commit (656f59d) seems to be the one where a new adjoint test (adjoint_test
) was introduced. The old adjoint_AAt_test
seems to pass on GPU.
The new test seems to be testing the adjoint accuracy in a different way than the old test.
I don't understand the tests well enough to comment on why one passes and the other fails. Any ideas?
from scico.
from scico.
That sounds like a reasonable solution to me.
It seems currently all the astra tests use the same tolerance irrespective of the test. I found that I can make the test pass by making the tolerance of adjoint_test
15 times. I have pushed the change in a (draft) pull request. Let me know your comments on the pull request.
from scico.
Related Issues (20)
- Setting up Sum of Weighted Convolutions Forward Operator HOT 3
- Deconvolution using Convolution Forward Operator with PDHG throwing _rmul_ not defined error HOT 9
- Custom Linear Operator with CircularConvolve Operation throwing error with ADMM solver HOT 4
- Re-open issue "Proximal operator for a custom prior #437" HOT 1
- Linear operator multiplication with singleton array not supported HOT 1
- Use of TV prior in Accelerated PGM solver HOT 4
- Estimating L0 parameter in AcceleratedPGM HOT 4
- Memory usage explosion when using AcceleratedPGM HOT 8
- Tuning penalty parameter in ProximalADMM for better deconvolution results HOT 2
- Enabling use of TV prior with a forward operator with Accelerated PGM HOT 2
- Reevaluate need for `ensure_on_device` function HOT 1
- Applying multiple priors with APGM solver HOT 7
- Solve AcceleratedPGM with multiple priors applied HOT 2
- Incorrect handling of scale in `Loss.grad` HOT 1
- Operator initialization failure
- Scaling of `FiniteDifference` linear operator broken
- Incorrect handling of functions with no return value in `scico.numpy`
- Potential issues in `scico.solver`
- Misleading extension on flax data files
- Updates to `scico.flax` needed due to changes in `orbax-checkpoint` API
Recommend Projects
-
React
A declarative, efficient, and flexible JavaScript library for building user interfaces.
-
Vue.js
🖖 Vue.js is a progressive, incrementally-adoptable JavaScript framework for building UI on the web.
-
Typescript
TypeScript is a superset of JavaScript that compiles to clean JavaScript output.
-
TensorFlow
An Open Source Machine Learning Framework for Everyone
-
Django
The Web framework for perfectionists with deadlines.
-
Laravel
A PHP framework for web artisans
-
D3
Bring data to life with SVG, Canvas and HTML. 📊📈🎉
-
Recommend Topics
-
javascript
JavaScript (JS) is a lightweight interpreted programming language with first-class functions.
-
web
Some thing interesting about web. New door for the world.
-
server
A server is a program made to process requests and deliver data to clients.
-
Machine learning
Machine learning is a way of modeling and interpreting data that allows a piece of software to respond intelligently.
-
Visualization
Some thing interesting about visualization, use data art
-
Game
Some thing interesting about game, make everyone happy.
Recommend Org
-
Facebook
We are working to build community through open source technology. NB: members must have two-factor auth.
-
Microsoft
Open source projects and samples from Microsoft.
-
Google
Google ❤️ Open Source for everyone.
-
Alibaba
Alibaba Open Source for everyone
-
D3
Data-Driven Documents codes.
-
Tencent
China tencent open source team.
from scico.